|
0981a08…
|
noreply
|
1 |
"""LiteLLM universal proxy provider implementation.""" |
|
0981a08…
|
noreply
|
2 |
|
|
0981a08…
|
noreply
|
3 |
import base64 |
|
0981a08…
|
noreply
|
4 |
import logging |
|
0981a08…
|
noreply
|
5 |
from pathlib import Path |
|
0981a08…
|
noreply
|
6 |
from typing import Optional |
|
0981a08…
|
noreply
|
7 |
|
|
0981a08…
|
noreply
|
8 |
from dotenv import load_dotenv |
|
0981a08…
|
noreply
|
9 |
|
|
0981a08…
|
noreply
|
10 |
from video_processor.providers.base import BaseProvider, ModelInfo, ProviderRegistry |
|
0981a08…
|
noreply
|
11 |
|
|
0981a08…
|
noreply
|
12 |
load_dotenv() |
|
0981a08…
|
noreply
|
13 |
logger = logging.getLogger(__name__) |
|
0981a08…
|
noreply
|
14 |
|
|
0981a08…
|
noreply
|
15 |
|
|
0981a08…
|
noreply
|
16 |
class LiteLLMProvider(BaseProvider): |
|
0981a08…
|
noreply
|
17 |
"""LiteLLM universal proxy provider. |
|
0981a08…
|
noreply
|
18 |
|
|
0981a08…
|
noreply
|
19 |
LiteLLM supports 100+ LLM providers through a unified interface. |
|
0981a08…
|
noreply
|
20 |
It reads provider API keys from environment variables automatically |
|
0981a08…
|
noreply
|
21 |
(e.g. OPENAI_API_KEY, ANTHROPIC_API_KEY, etc.). |
|
0981a08…
|
noreply
|
22 |
""" |
|
0981a08…
|
noreply
|
23 |
|
|
0981a08…
|
noreply
|
24 |
provider_name = "litellm" |
|
0981a08…
|
noreply
|
25 |
|
|
0981a08…
|
noreply
|
26 |
def __init__(self): |
|
0981a08…
|
noreply
|
27 |
try: |
|
0981a08…
|
noreply
|
28 |
import litellm # noqa: F401 |
|
0981a08…
|
noreply
|
29 |
except ImportError: |
|
0981a08…
|
noreply
|
30 |
raise ImportError("litellm package not installed. Install with: pip install litellm") |
|
0981a08…
|
noreply
|
31 |
|
|
0981a08…
|
noreply
|
32 |
self._litellm = litellm |
|
0981a08…
|
noreply
|
33 |
self._last_usage = {} |
|
0981a08…
|
noreply
|
34 |
|
|
0981a08…
|
noreply
|
35 |
def chat( |
|
0981a08…
|
noreply
|
36 |
self, |
|
0981a08…
|
noreply
|
37 |
messages: list[dict], |
|
0981a08…
|
noreply
|
38 |
max_tokens: int = 4096, |
|
0981a08…
|
noreply
|
39 |
temperature: float = 0.7, |
|
0981a08…
|
noreply
|
40 |
model: Optional[str] = None, |
|
0981a08…
|
noreply
|
41 |
) -> str: |
|
0981a08…
|
noreply
|
42 |
if not model: |
|
0981a08…
|
noreply
|
43 |
raise ValueError( |
|
0981a08…
|
noreply
|
44 |
"LiteLLM requires an explicit model in provider/model format " |
|
0981a08…
|
noreply
|
45 |
"(e.g. 'openai/gpt-4o', 'anthropic/claude-3-sonnet-20240229')" |
|
0981a08…
|
noreply
|
46 |
) |
|
0981a08…
|
noreply
|
47 |
|
|
0981a08…
|
noreply
|
48 |
response = self._litellm.completion( |
|
0981a08…
|
noreply
|
49 |
model=model, |
|
0981a08…
|
noreply
|
50 |
messages=messages, |
|
0981a08…
|
noreply
|
51 |
max_tokens=max_tokens, |
|
0981a08…
|
noreply
|
52 |
temperature=temperature, |
|
0981a08…
|
noreply
|
53 |
) |
|
0981a08…
|
noreply
|
54 |
|
|
0981a08…
|
noreply
|
55 |
usage = getattr(response, "usage", None) |
|
0981a08…
|
noreply
|
56 |
self._last_usage = { |
|
0981a08…
|
noreply
|
57 |
"input_tokens": getattr(usage, "prompt_tokens", 0) if usage else 0, |
|
0981a08…
|
noreply
|
58 |
"output_tokens": getattr(usage, "completion_tokens", 0) if usage else 0, |
|
0981a08…
|
noreply
|
59 |
} |
|
0981a08…
|
noreply
|
60 |
return response.choices[0].message.content or "" |
|
0981a08…
|
noreply
|
61 |
|
|
0981a08…
|
noreply
|
62 |
def analyze_image( |
|
0981a08…
|
noreply
|
63 |
self, |
|
0981a08…
|
noreply
|
64 |
image_bytes: bytes, |
|
0981a08…
|
noreply
|
65 |
prompt: str, |
|
0981a08…
|
noreply
|
66 |
max_tokens: int = 4096, |
|
0981a08…
|
noreply
|
67 |
model: Optional[str] = None, |
|
0981a08…
|
noreply
|
68 |
) -> str: |
|
0981a08…
|
noreply
|
69 |
if not model: |
|
0981a08…
|
noreply
|
70 |
raise ValueError( |
|
0981a08…
|
noreply
|
71 |
"LiteLLM requires an explicit model for image analysis " |
|
0981a08…
|
noreply
|
72 |
"(e.g. 'openai/gpt-4o', 'anthropic/claude-3-sonnet-20240229')" |
|
0981a08…
|
noreply
|
73 |
) |
|
0981a08…
|
noreply
|
74 |
|
|
0981a08…
|
noreply
|
75 |
b64 = base64.b64encode(image_bytes).decode() |
|
0981a08…
|
noreply
|
76 |
|
|
0981a08…
|
noreply
|
77 |
response = self._litellm.completion( |
|
0981a08…
|
noreply
|
78 |
model=model, |
|
0981a08…
|
noreply
|
79 |
messages=[ |
|
0981a08…
|
noreply
|
80 |
{ |
|
0981a08…
|
noreply
|
81 |
"role": "user", |
|
0981a08…
|
noreply
|
82 |
"content": [ |
|
0981a08…
|
noreply
|
83 |
{"type": "text", "text": prompt}, |
|
0981a08…
|
noreply
|
84 |
{ |
|
0981a08…
|
noreply
|
85 |
"type": "image_url", |
|
0981a08…
|
noreply
|
86 |
"image_url": {"url": f"data:image/jpeg;base64,{b64}"}, |
|
0981a08…
|
noreply
|
87 |
}, |
|
0981a08…
|
noreply
|
88 |
], |
|
0981a08…
|
noreply
|
89 |
} |
|
0981a08…
|
noreply
|
90 |
], |
|
0981a08…
|
noreply
|
91 |
max_tokens=max_tokens, |
|
0981a08…
|
noreply
|
92 |
) |
|
0981a08…
|
noreply
|
93 |
|
|
0981a08…
|
noreply
|
94 |
usage = getattr(response, "usage", None) |
|
0981a08…
|
noreply
|
95 |
self._last_usage = { |
|
0981a08…
|
noreply
|
96 |
"input_tokens": getattr(usage, "prompt_tokens", 0) if usage else 0, |
|
0981a08…
|
noreply
|
97 |
"output_tokens": getattr(usage, "completion_tokens", 0) if usage else 0, |
|
0981a08…
|
noreply
|
98 |
} |
|
0981a08…
|
noreply
|
99 |
return response.choices[0].message.content or "" |
|
0981a08…
|
noreply
|
100 |
|
|
0981a08…
|
noreply
|
101 |
def transcribe_audio( |
|
0981a08…
|
noreply
|
102 |
self, |
|
0981a08…
|
noreply
|
103 |
audio_path: str | Path, |
|
0981a08…
|
noreply
|
104 |
language: Optional[str] = None, |
|
0981a08…
|
noreply
|
105 |
model: Optional[str] = None, |
|
0981a08…
|
noreply
|
106 |
) -> dict: |
|
0981a08…
|
noreply
|
107 |
model = model or "whisper-1" |
|
0981a08…
|
noreply
|
108 |
|
|
0981a08…
|
noreply
|
109 |
try: |
|
0981a08…
|
noreply
|
110 |
with open(audio_path, "rb") as f: |
|
0981a08…
|
noreply
|
111 |
response = self._litellm.transcription( |
|
0981a08…
|
noreply
|
112 |
model=model, |
|
0981a08…
|
noreply
|
113 |
file=f, |
|
0981a08…
|
noreply
|
114 |
language=language, |
|
0981a08…
|
noreply
|
115 |
) |
|
0981a08…
|
noreply
|
116 |
|
|
0981a08…
|
noreply
|
117 |
text = getattr(response, "text", str(response)) |
|
0981a08…
|
noreply
|
118 |
self._last_usage = { |
|
0981a08…
|
noreply
|
119 |
"input_tokens": 0, |
|
0981a08…
|
noreply
|
120 |
"output_tokens": 0, |
|
0981a08…
|
noreply
|
121 |
} |
|
0981a08…
|
noreply
|
122 |
|
|
0981a08…
|
noreply
|
123 |
return { |
|
0981a08…
|
noreply
|
124 |
"text": text, |
|
0981a08…
|
noreply
|
125 |
"segments": [], |
|
0981a08…
|
noreply
|
126 |
"language": language, |
|
0981a08…
|
noreply
|
127 |
"duration": None, |
|
0981a08…
|
noreply
|
128 |
"provider": "litellm", |
|
0981a08…
|
noreply
|
129 |
"model": model, |
|
0981a08…
|
noreply
|
130 |
} |
|
0981a08…
|
noreply
|
131 |
except Exception: |
|
0981a08…
|
noreply
|
132 |
raise NotImplementedError( |
|
0981a08…
|
noreply
|
133 |
"Audio transcription failed via LiteLLM. " |
|
0981a08…
|
noreply
|
134 |
"Ensure the underlying provider supports transcription." |
|
0981a08…
|
noreply
|
135 |
) |
|
0981a08…
|
noreply
|
136 |
|
|
0981a08…
|
noreply
|
137 |
def list_models(self) -> list[ModelInfo]: |
|
0981a08…
|
noreply
|
138 |
try: |
|
0981a08…
|
noreply
|
139 |
model_list = getattr(self._litellm, "model_list", None) |
|
0981a08…
|
noreply
|
140 |
if model_list: |
|
0981a08…
|
noreply
|
141 |
return [ |
|
0981a08…
|
noreply
|
142 |
ModelInfo( |
|
0981a08…
|
noreply
|
143 |
id=m if isinstance(m, str) else str(m), |
|
0981a08…
|
noreply
|
144 |
provider="litellm", |
|
0981a08…
|
noreply
|
145 |
display_name=m if isinstance(m, str) else str(m), |
|
0981a08…
|
noreply
|
146 |
capabilities=["chat"], |
|
0981a08…
|
noreply
|
147 |
) |
|
0981a08…
|
noreply
|
148 |
for m in model_list |
|
0981a08…
|
noreply
|
149 |
] |
|
0981a08…
|
noreply
|
150 |
except Exception as e: |
|
0981a08…
|
noreply
|
151 |
logger.warning(f"Failed to list LiteLLM models: {e}") |
|
0981a08…
|
noreply
|
152 |
return [] |
|
0981a08…
|
noreply
|
153 |
|
|
0981a08…
|
noreply
|
154 |
|
|
0981a08…
|
noreply
|
155 |
# Only register if litellm is importable |
|
0981a08…
|
noreply
|
156 |
try: |
|
0981a08…
|
noreply
|
157 |
import litellm # noqa: F401 |
|
0981a08…
|
noreply
|
158 |
|
|
0981a08…
|
noreply
|
159 |
ProviderRegistry.register( |
|
0981a08…
|
noreply
|
160 |
name="litellm", |
|
0981a08…
|
noreply
|
161 |
provider_class=LiteLLMProvider, |
|
0981a08…
|
noreply
|
162 |
env_var="", |
|
0981a08…
|
noreply
|
163 |
model_prefixes=[], |
|
0981a08…
|
noreply
|
164 |
default_models={ |
|
0981a08…
|
noreply
|
165 |
"chat": "", |
|
0981a08…
|
noreply
|
166 |
"vision": "", |
|
0981a08…
|
noreply
|
167 |
"audio": "", |
|
0981a08…
|
noreply
|
168 |
}, |
|
0981a08…
|
noreply
|
169 |
) |
|
0981a08…
|
noreply
|
170 |
except ImportError: |
|
0981a08…
|
noreply
|
171 |
pass |